AI
“Microsoft just unveiled six in-house AI models to compete with OpenAI and Anthropic on performance and pricing.”
Plain restatementMicrosoft announced a set of new first-party AI models, positioned against OpenAI and Anthropic offerings on both capability and cost.
Microsoft really did announce a batch of new in-house AI models at its Ignite conference on November 17, 2026, and the figures quoted in this post match Microsoft's own published wording. Two things are off. First, Microsoft's own announcement describes seven models, not six, with a decision-making model left out of the post's list, even though the post credits Microsoft as its source. Second, and more important, the eye-catching price comparisons such as "40% of the price" and "20% of the cost" were calculated using a 50% promotional discount that expires in May 2027. At the regular price the larger coding model costs about 81% of the rival it is compared to rather than 40%, and the smaller model actually ends up about 30% more expensive than the OpenAI model it is said to undercut. Every performance and cost comparison here was run by Microsoft on benchmarks Microsoft chose, and no independent testing of these coding or security models has been published yet. The image editing top ranking is real on one specific leaderboard, but the lead there is small enough that the board itself treats the model as tied with two rivals. The post does say "according to Microsoft" in most places, which is fair, but a reader would still come away thinking these prices are the prices they would pay.
Microsoft [drifted from the evidence:] just unveiled six in-house AI models [drifted from the evidence:] to compete with OpenAI and Anthropic on [drifted from the evidence:] performance and [drifted from the evidence:] pricing.
Microsoft [added by the neutral restatement:] announced a set of new first-party AI models, [added by the neutral restatement:] positioned against OpenAI and Anthropic [added by the neutral restatement:] offerings on [added by the neutral restatement:] both capability and [added by the neutral restatement:] cost.
Red-tinted words in the claim drifted from the evidence. Green-tinted words are what a neutral restatement needs.
The trace / claim to source
- Microsoft did unveil a family of new in-house AI models at Ignite on November 17, 2026, explicitly positioned against OpenAI and Anthropic on both capability and cost.
- The MAI Code 2 figures quoted in the post match Microsoft's own keynote wording exactly: 98% of GPT 6.1 Sol's performance at 40% of the cost, and 96% of Opus 5.5's performance at 5% of the cost, averaged over four benchmarks.
- The MAI Code 2 Flash figures quoted match Microsoft's announcement: matching GPT-6 Luna at 65% of the price and Haiku 5.5 at 20% of the price.
- The cybersecurity line matches the keynote, which places MAI Cyber 2 Flash ahead of GPT 6 Astra on Cyber Gym end-to-end at less than 2% of the cost per task.
- MAI Code 2 is in fact live in GitHub Copilot, generally available across Pro, Pro+, Max, Business and Enterprise plans.
- The enterprise weights point is accurate to the vendor's own description of MAI Licensed Weights, under which a customer customizes a MAI model with proprietary data and runs it in its own Azure environment.
- The Suleyman point is accurately reported: he told Axios Microsoft is making significant progress toward its own frontier general-purpose model, and no release date was given.
- MAI Image 2.7 does hold the top point score on the Artificial Analysis image editing board, and the post correctly attributes the claim to Microsoft.
- The post attributes most of its performance figures to Microsoft rather than presenting them as independent results, which is accurate sourcing practice.
- The post presents "40% of the price," "65% of the price" and "20% of the cost" as the models' prices, with no mention that Microsoft's cost comparisons were computed at a 50% promotional rate that runs only until May 17, 2027. At the standing list price the same per-task arithmetic puts MAI Code 2 at roughly 81% of GPT-6.1 Sol's cost rather than 40%, and puts MAI Code 2 Flash roughly 30% above GPT-6 Luna rather than at 65% of its price. The headline claim names pricing as one of its two axes, so this changes the meaning of half the claim. Microsoft itself footnoted the promotional basis.
- The coding comparison is scoped by Microsoft to everyday enterprise coding tasks of the kind seen in GitHub Copilot, and Microsoft separately said the coding models are not tuned for harder work such as basic science research. The post carries the numbers without the scope, which invites a reader to take them as general coding superiority.
- The post says six models, while Microsoft's own announcement and keynote describe seven. The missing model is Microsoft-Decision-1.1, which is generally available in Foundry. The post credits "Source: Microsoft" for a count that does not match Microsoft's post. The number six traces to Axios, which counted six; one outlet noted the discrepancy plainly, writing that Microsoft's own post describes seven new models while Axios, which interviewed Suleyman, counted six.
- "ranks first in image editing" is true on one specific board, AA-Image-Editing v2.1, and the post gives no board name. On that board's own ranking rule the model sits in a 1-3 rank band, statistically tied with two OpenAI models whose point scores are 4 and 10 Elo lower. Microsoft's own earlier preview post for the same model reported it at No. 3 on Arena Edit and No. 2 on Artificial Analysis's image editing leaderboard, so the ranking depends on which board and which day is cited.
- The post lists the models as a single unveiling without distinguishing stages. In reality MAI Code 2 and Flash are generally available in GitHub, Code 2 Flash is only in private preview in Foundry, and MAI Cyber 2 Flash is in preview and restricted to verified defenders.
- Whether the performance comparisons hold up outside Microsoft's own harness. No independent runner has published results on MAI Code 2, MAI Code 2 Flash or MAI Cyber 2 Flash against the named competitors as of November 18, 2026. Every coding and security number in the post is a vendor self-report on vendor-chosen benchmarks.
- Whether competitor cost figures used in Microsoft's per-task tables reflect each rival's current list pricing, and whether those rivals' prices move before the promotional period ends.
- Whether the image editing top rank persists. The lead is 4 Elo inside a ±14 interval on a vote-accumulating board, so the ordering can change without any model changing.
- The eligibility criteria, pricing and scope of the MAI Licensed Weights program, which Microsoft describes in general terms and gates to selected enterprise customers.
- Whether "six" versus "seven" reflects a judgment about what counts as a model, for example whether Dragon Copilot Transcribe is counted as a product or a model, or a simple undercount.
The underlying event is real and occurred on the date claimed. Microsoft AI's own announcement describes a family of seven new models at "scalable price-performance," listing MAI Realtime 1 as its first live speech-to-speech model, said to outperform GPT-Realtime 2.1 across external and Dynamics 365 benchmarks, and Dragon Copilot Transcribe built on MAI Transcribe 2. Reporting lists the lineup as MAI Cyber 2 Flash, MAI Code 2, MAI Code 2 Flash, MAI Image 2.7, MAI Realtime 1, Dragon Copilot Transcribe and Decision 1.1, announced November 17 at Ignite. On the headline coding numbers, Microsoft's keynote states them directly: averaged over four widely used benchmarks targeting everyday enterprise coding tasks seen in GitHub Copilot, MAI Code 2 gives 98% of the performance of GPT 6.1 Sol at 40% of the cost, and against Opus 5.5 it is 5% of that cost. The same transcript covers the security model: MAI Cyber 2 Flash is described as ahead of GPT 6 Astra on Cyber Gym end-to-end at less than 2% of the cost per task. The company announcement also carries the Flash and image lines: MAI Code 2 Flash matches the performance of GPT-6 Luna at 65% of the price and Haiku 5.5 at 20% of the price, and MAI Image 2.7 is #1 on Artificial Analysis image editing, outperforming GPT Image 2.6 Sunburst (max) at one fifth of the price. The image ranking checks out against the live board, with a caveat. Microsoft states that as of November 16, 2026, MAI Image 2.7 ranks #1 on AA-Image-Editing v2.1, scoring 1,215 Elo across 5,478 votes, ahead of GPT Image 2.6 Sunburst (max) by 4 Elo. The leaderboard itself confirms the point score and shows the ranking rule: MAI-Image-2.7 leads with Elo 1215, listed at rank 1 with a rank range of 1-3 and a 95% interval of 1201 to 1229, ahead of GPT Image 2.6 Sunburst (max) at 1211 and GPT Image 2.6 Flare (max) at 1205. The enterprise access point is accurate to the vendor's wording: the keynote describes MAI Licensed Weights, where a customer takes a frontier MAI model, customizes it using proprietary data, and runs it securely in its own Azure environment. The Suleyman statement is accurately reported: Axios reports that Suleyman told it Microsoft is making significant progress toward building its own frontier general-purpose AI model, even as it introduces specialized models that it claims rival OpenAI and Anthropic offerings at a fraction of the cost. The decisive omitted condition is the pricing basis. The cost figures used for comparison are at 50% promotional pricing, with MAI Code 2 available at $0.50 per million input tokens and $2.50 per million output tokens until May 17, 2027 according to GitHub's pricing page. At the $1/$5 list price, the same tasks cost about 81% of what Sol charges. For the smaller model the comparison inverts at list price: MAI Code 2 Flash at $0.052 per task beats GPT-6 Luna's $0.080 today, but the doubled figure of $0.104 sits about 30% above it, assuming the rival's cost does not move.
Complete reasoning
The reply is formatted for pasting into the thread where the claim is circulating.
Ask this case
Answers come only from the case file above; nothing is added.
Did Microsoft really announce new AI models to compete with OpenAI and Anthropic?
Yes. Microsoft announced a family of new in-house AI models at its Ignite conference on November 17, 2026, explicitly positioned against OpenAI and Anthropic on capability and cost.
Is it true that Microsoft unveiled six models?
No. Microsoft's own announcement and keynote describe seven models, including a decision-making model called Decision 1.1 that the post left out. The number six appears to come from a different outlet's count, not Microsoft's own post, even though the post credits Microsoft as its source.
Are the price comparisons like '40% of the price' accurate?
They match Microsoft's own figures, but those figures were calculated using a 50% promotional discount that expires in May 2027. At the regular list price, the larger coding model costs about 81% of the rival's price rather than 40%, and the smaller model ends up about 30% more expensive than the OpenAI model it is compared to.
Is MAI Image 2.7 really ranked number one for image editing?
It does hold the top point score on one specific leaderboard, AA-Image-Editing v2.1, as Microsoft stated. However, the lead is small enough that the leaderboard itself places it in a rank band of 1 to 3, meaning it is statistically tied with two other models rather than clearly in first place.
Have these performance and cost claims been independently verified?
The case file does not establish that. Every performance and cost comparison was run by Microsoft on benchmarks Microsoft chose, and no independent testing of the coding or security models has been published yet.